U-PINet: End-to-End Hierarchical Physics-Informed Learning With Sparse Graph Coupling for 3D EM Scattering Modeling

Rui Zhu, Yuexing Peng, Peng Wang, George C. Alexandropoulos, Wenbo Wang, Wei Xiang

Abstract

Accurate radar cross section (RCS) computation is a fundamental task in radar engineering and electromagnetic (EM) scattering analysis, underpinning target signature characterization, detection, and recognition. Conventional computational electromagnetics (CEM) solvers provide high-fidelity RCS predictions but suffer from prohibitive computational costs when applied to 3-dimensional (3D) targets under multi-aspect configurations. In contrast, purely data-driven neural networks offer high efficiency yet often lack physical consistency and generalization capability. To address these challenges, this paper proposes a U-shaped Physics-Informed Network (U-PINet). To the best of our knowledge, it is the first framework to establish a fully end-to-end, physics-informed hierarchical architecture for fast and accurate RCS computation, grounded in the governing principles of CEM. Inspired by the near-far field decomposition in classical fast solvers, U-PINet explicitly models local EM coupling and long-range radiation effects through a hierarchical operator design. A physics-guided graph construction is further introduced to represent self- and mutual-coupling among mesh elements of complex 3D targets, enabling physically interpretable intermediate representations. By embedding EM governing equations as residual constraints, the proposed framework achieves end-to-end, physically consistent RCS prediction with significantly improved computational efficiency. Extensive numerical experiments demonstrate that U-PINet attains solver-level RCS accuracy with orders-of-magnitude runtime reduction, while exhibiting strong generalization to unseen target geometries under limited training data.

Introduction

Predicting how an object scatters radar waves—its Radar Cross Section (RCS)—is critical for applications ranging from stealth technology to autonomous vehicle sensing. Traditionally, this is achieved using high-fidelity Computational Electromagnetics (CEM) solvers like the Method of Moments (MoM) or the Multilevel Fast Multipole Algorithm (MLFMA). While accurate, these methods are computationally intensive, often requiring hours or days to simulate complex 3D targets because they involve solving massive, dense systems of linear equations.

Recent deep learning approaches offer a faster alternative, but they typically treat the problem as a "black box," regressing RCS directly from geometry. These purely data-driven models often lack physical consistency and fail to generalize to unseen shapes or orientations. U-PINet bridges this gap by creating a model that is as fast as a neural network but as physically grounded as a traditional solver.

Methodology: The Neural Twin of Classical Solvers

U-PINet's architecture is designed as a "neural twin" of the hierarchical MLFMA solver. It mirrors the structure of classical fast solvers by explicitly separating electromagnetic interactions into near-field and far-field components.

Comparison between traditional MLFMA and the proposed U-PINet framework.

As shown in the comparison above, U-PINet replaces the iterative numerical calculations of traditional solvers with a learnable hierarchical architecture. This design allows the network to capture the same physical phenomena—local coupling and global radiation—while benefiting from the inference speed of deep learning.

Architectural Components

The framework consists of two primary modules that handle different scales of electromagnetic interaction:

  1. Near-Field Graph Encoder: To model local electromagnetic coupling, U-PINet employs a Physics-Guided Graph Neural Network (GNN). It treats mesh elements as nodes and utilizes a custom edge kernel that accounts for inter-point distance and surface normal alignment. This mimics the Green’s function used in physics to represent self- and mutual-coupling among adjacent mesh elements.
  2. Hierarchical Far-Field Fusion: To handle long-range radiation, the network uses a multi-scale U-shaped structure. It aggregates local features into coarser representations (fine-to-coarse) and uses cross-scale attention mechanisms to model how distant parts of an object interact, effectively performing the "aggregation, translation, and disaggregation" steps found in MLFMA.
Detailed architecture of U-PINet showing the Near-Field Graph Encoder and Hierarchical Far-Field Fusion modules.

Physics-Informed Learning

Unlike standard AI models that only minimize the difference between predicted and true labels, U-PINet minimizes the residual of the Electric Field Integral Equation (EFIE). By embedding Maxwell’s equations directly into the loss function, the network is forced to ensure that the predicted surface currents ($J_{pred}$) are physically consistent.

Instead of regressing the final RCS value directly, U-PINet reconstructs the underlying 3D surface currents. The final RCS ($\sigma$) is then calculated via a standard radiation integral. This approach ensures that the output is a mathematical consequence of physical currents, making the model significantly more robust and generalizable than direct regression models.

Key Results and Performance

The researchers evaluated U-PINet using several 3D Perfect Electrically Conducting (PEC) targets, including canonical shapes (spheres, cubes), complex benchmarks (the SLICY model), and large-scale ship models. Ground truth data was generated using high-fidelity MLFMA simulations at 1 GHz.

  • Efficiency: For a complex ship model, a traditional MoM solver required over 80,000 seconds (~22 hours). U-PINet completed the same task in just 0.12 seconds, representing an orders-of-magnitude speedup.
  • Accuracy: U-PINet achieved "solver-level" accuracy, with a Coefficient of Determination ($R^2$) exceeding 0.97 for simple objects and maintaining high fidelity on intricate targets like the SLICY model.
  • Generalization: In "leave-one-geometry-out" tests, the model maintained high accuracy even when trained on only 10% of the data for a specific complex shape, significantly outperforming purely data-driven CNNs and standard U-Nets.

Conclusion

U-PINet represents a significant advancement in EM scattering modeling by successfully integrating the structural logic of classical CEM solvers with the efficiency of deep learning. By focusing on the reconstruction of physical surface currents rather than black-box regression, it provides a physically consistent and highly efficient surrogate for traditional solvers. Future work will explore extending the framework to broadband frequency analysis and applying it to downstream tasks such as Synthetic Aperture Radar (SAR) imaging.

Summary updated

Abstract

Accurate radar cross section (RCS) computation is a fundamental task in radar engineering and electromagnetic (EM) scattering analysis, underpinning target signature characterization, detection, and recognition. Conventional computational electromagnetics (CEM) solvers provide high-fidelity RCS predictions but suffer from prohibitive computational costs when applied to 3-dimensional (3D) targets under multi-aspect configurations. In contrast, purely data-driven neural networks offer high efficiency yet often lack physical consistency and generalization capability. To address these challenges, this paper proposes a U-shaped Physics-Informed Network (U-PINet). To the best of our knowledge, it is the first framework to establish a fully end-to-end, physics-informed hierarchical architecture for fast and accurate RCS computation, grounded in the governing principles of CEM. Inspired by the near-far field decomposition in classical fast solvers, U-PINet explicitly models local EM coupling and long-range radiation effects through a hierarchical operator design. A physics-guided graph construction is further introduced to represent self- and mutual-coupling among mesh elements of complex 3D targets, enabling physically interpretable intermediate representations. By embedding EM governing equations as residual constraints, the proposed framework achieves end-to-end, physically consistent RCS prediction with significantly improved computational efficiency. Extensive numerical experiments demonstrate that U-PINet attains solver-level RCS accuracy with orders-of-magnitude runtime reduction, while exhibiting strong generalization to unseen target geometries under limited training data.
BibTeX
@misc{zhu2025u,
  author = {Rui Zhu and Yuexing Peng and Peng Wang and George C. Alexandropoulos and Wenbo Wang and Wei Xiang},
  title = {{U-PINet: End-to-End Hierarchical Physics-Informed Learning With Sparse Graph Coupling for 3D EM Scattering Modeling}},
  year = {2025},
  eprint = {2508.03774},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.03774},
}
Figure